To build a detailed knowledge of the biodiversity, the geo- graphical distribution and the evolution of the alive species is essential for a sustainable development and the preser- vation of this biodiversity. Massive databases of underwa- ter video surveillance have been recently made available for supporting designing algorithms targeting the identification of fishes. However these video datasets are rather poor in terms of video resolution, pretty challenging regarding both the natural phenomena to be considered such as murky wa- ter, seaweed moving the water current, etc, and the huge amount of data to be processed. We have designed a processing chain based on background segmentation, selection keypoints with an adaptive scale, description with OpponentSift and learning of each species by a binary linear Support Vector Machines classifier. Our algorithm has been evaluated in the context of our participation to the Fish task of the LifeCLEF2014 chal- lenge. Compared to the baseline designed by the LifeCLEF challenge organizers, our approach reaches a better precision but a worse recall. Our performances in terms of species recognition (based only on the correctly detected bounding boxes) is comparable to the baseline, but our bounding boxes are often too large and our score is so penalized. Our results are thus really encouraging.